convez.ai
MCP ServerBusca & PesquisaAvançado
jaccen

Awesome Gaussian Skills

jaccen/Awesome-Gaussian-Skills

660+ métodos 3DGS com skills de IA e análise de bugs para pesquisa gráfica

Ver no GitHub
BomPluginScore · de 0 a 100

Catálogo abrangente com 660+ métodos de 3D Gaussian Splatting, NeRF e computação gráfica, organizados em 25 categorias. Inclui 12 skills de IA e 97+ padrões de bugs para pesquisa e desenvolvimento em visão 3D, com explorador interativo e compatibilidade com OpenClaw/Claude Code.

Instalação

$npx convez-ai jaccen/Awesome-Gaussian-Skills

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
188
Forks
13
Acompanhando
3
Saúde
57%
Último commit
há 2 dias
Versão
v0.9.8
Issues abertas
1
Contribuidores
3

Casos de uso

  • Pesquisa em 3D Gaussian Splatting e NeRF
  • Análise e catalogação de métodos de reconstrução 3D
  • Detecção de padrões de bugs em código gráfico 3D
  • Exploração interativa de papers e técnicas de síntese de cenas

Veredito

Recurso de referência sólido para pesquisadores em visão 3D, com abrangência impressionante (660+ métodos) e skills de IA que agregam valor além de simples listagem. Execução técnica madura, mas requer familiaridade com 3DGS/NeRF para aproveitar plenamente.

README

A tradução para o português deste README ainda está na fila.

Awesome Gaussian Skills

The Most Comprehensive 3DGS Paper Catalog — 660+ Methods, 25 Categories, Interactive Explorer

You shouldn't search 20 different repos for 3DGS papers. This is the only one you need — plus AI-powered tools that no other list has.

Stars Live Demo Methods Skills Bug Patterns License PRs Welcome

English | 中文

What's New (June 2026)

Latest update (Jun 5): 630+ → 660+ methods, 25 categories, 97+ bug patterns. +18 new methods including: ZipSplat (token-based feed-forward 3DGS, ~6x fewer Gaussians, +2.1dB PSNR), Geometry Gaussians (per-splat geometry opacity decoupling), 3DReflecNet (CVPR 2026, systematic benchmark for reflective/transparent 3D reconstruction), MeshFlow (CVPR 2026 Highlight, MeshVAE + Flow-based Diffusion Transformer), MeshWeaver (CVPR 2026, autoregressive next-vertex mesh generation), UniCAD (multi-modal multi-task CAD benchmark), Anchor3R (streaming 3D reconstruction with transient anchors), SimuScene (physics-in-the-loop compositional 3D reconstruction). Skills v0.3.1: 5 new bug patterns (#89-#93) + Token-based cluster in method comparison.

MethodArXivCategoryOne-Line Innovation
D4RT2512.08924Dynamic / 4DCVPR 2026 Best Paper; unified query 4D reconstruction + tracking; 200+ FPS
TRELLIS.2—GenerationCVPR 2026 Best Student Paper; 4B-param native 3D; 17s PBR generation
SAM 3D—Generation / FoundationCVPR 2026 Honorable Mention; single-image 3D foundation model (Meta)
RAF—Physics / SimulationCVPR 2026 Findings; bridges 3DGS and physics engines (MPM/SPH/PBD)
ReLaGS2603.17605Language / SemanticOpen-vocab 3D reasoning without per-scene training
FreeForm2605.29318Dynamic / ElasticParticle-skinned eigenmodes for elastic deformation on 3DGS
FreeArtGS2603.22102ArticulatedArticulated GS under free-moving scenario
PARTICULATE—Articulated / Feed-ForwardFeed-forward 3D object articulation from static mesh
DropAnSH-GS—Sparse-ViewAnchor dropout + SH regularization for sparse-view 3DGS
BA-GS—SfM-FreeBayesian Adaptive GS for SfM-free reconstruction
SR3R—Feed-Forward / SRSuper-resolution + feed-forward GS reconstruction
ZipSplat2606.05102Feed-ForwardToken-based 3DGS; ~6x fewer Gaussians; +2.1dB PSNR over pixel-grid methods
Geometry Gaussians2606.05124Surface / RenderingPer-splat geometry opacity decoupling; signed-opacity direction
3DReflecNet2605.10204Cross-DomainCVPR 2026; systematic benchmark for reflective/transparent 3D reconstruction
MeshFlow2606.04621CAD / GenerationCVPR 2026 Highlight; MeshVAE + Flow-based Diffusion Transformer
MeshWeaver2606.04688CAD / GenerationCVPR 2026; autoregressive next-vertex mesh generation
UniCAD2606.05058CADMulti-modal multi-task CAD benchmark + universal model
Anchor3R2606.05035Feed-ForwardStreaming 3D reconstruction with transient anchors
SimuScene2606.03994SimulationPhysics-in-the-loop compositional 3D reconstruction
HiGS2606.00352AccelerationHierarchical tiles decouple partitioning/rasterization; 15.8x faster (NVIDIA)
DDF-GS2606.00817Rendering / GIDirected Distance Field for ray queries on 3DGS; GI without mesh proxy
VEDAL2606.02346CompressionVariational free energy pruning; 5.2x compression @ 0.31 dB loss
3DThinkVLA2606.04436RoboticsVLA with latent 3D priors via co-training
SymTRELLIS2606.04108GenerationSymmetry-aware 3D generation

Full changelog: changelog/

Knowledge base curated with multi-source verification. Found an error? Open an Issue.

Why This Project?

You shouldn't have to rebuild the same RAG pipeline for every new 3DGS paper, or manually compare 20 variants across 10 dimensions, or discover after submission that your CUDA kernel has a known bug.

With 660+ 3DGS papers since 2023 and growing, researchers waste hours on tasks that AI Agents can handle in seconds. Yet ClawHub's 13,000+ skills cover almost zero for 3D reconstruction or computer graphics.

What You DoWithout ThisWith This
Read a new paper30–60 min manual skimmingStructured summary in seconds
Compare GS variantsHand-build comparison tables10+ dimension auto-comparison
Review 3DGS codeMiss known bug patterns97+ pattern detection
Design experimentsGuess baselines & ablationsVenue-tailored experiment plan
Migrate NeRF → 3DGSTrial-and-error portingStep-by-step migration guide

How We Compare

What You NeedTypical Paper ListAwesome Gaussian Skills
Browse papersStatic markdown tableInteractive explorer: search, filter, sort
Compare methodsOpen 2 papers side by side10+ dimension auto-comparison
Avoid code bugsDiscover after submission97+ known bug pattern detection
Design experimentsGuess baselines & ablationsVenue-tailored experiment plan
NeRF → 3DGSTrial-and-error portingStep-by-step migration guide
CAD ↔ 3DGSNo coverage52+ method conversion pipeline
Patent filingManual from scratchAuto-generated claims & specs

Bottom line: Other lists give you paper titles. We give you paper titles + an AI toolkit that actually helps you do research faster.

Features

  • 660+ Methods Knowledge Base: The most comprehensive catalog of 3D Gaussian Splatting variants across 25 categories, with arXiv IDs, venues, innovations, and code links. Updated daily.
  • Interactive Explorer: Try it now — Search, filter by category, sort by citations, click any method card for details
  • 12 AI-Powered Skills (Advanced): Paper reading, method comparison, code review, spatial intelligence, MCP rendering, signed Gaussian splatting — works with OpenClaw, Claude Code, Cursor
  • Zero Setup: Pure Markdown files, no dependencies
  • Actively Maintained: Daily arXiv tracking

Knowledge Base (660+ Methods)

GroupCategoriesKey Topics
Core RepresentationsFoundation, Antialiasing, Optimization, Surface/Rendering, Image Rep.3DGS, 2DGS, Scaffold-GS, Mip-Splatting, GaussianImage...
Efficiency & ScaleCompression, Acceleration, Large-Scale, Feed-ForwardCompact-3DGS, BlitzGS, HiGS, VEDAL, VG²GT...
Understanding & SemanticsLanguage/Semantic, Generation, Autonomous DrivingLangSplat, DreamGaussian, StreetNVS...
Dynamic & SpatialDynamic, HDR, SLAM, Sparse-ViewDSD-GS, WebSpline, GGD-SLAM, PanoPlane...
ApplicationsHuman/Avatar, Editing, Relighting, CAD, Cross-Domain, Simulation, Robotics, and 14 more categoriesAlbedoEdit, KDH-CAD, LEGS, TIDES, 3DEditSafe...

Try the interactive method explorer — search 660+ methods instantly, filter by category, sort by citations.

Download full database: 3dgs-methods-overview.csv | Full analysis: references/3dgs-methods-overview.md

Core Representations

CategoryDescriptionMethods
FoundationCore 3DGS representations and basic variants3DGS, 2DGS, Scaffold-GS, Scaffold-GS+, Mip-Splatting, 3DGEER, SNS
AntialiasingAnti-aliasing and frequency-aware renderingMip-Splatting, LeanGaussian
OptimizationTraining objectives, density control, and convergence3DGS-as-MCMC, 3DGS², AdpSplit, Denoising-GS
Surface / RenderingRendering formulation innovation (OIT, RBF, etc.)SparseOIT, View-Dependent Splatting, Gaussian Surfel Rendering
Image RepresentationImage-level Gaussian encodingGaussianImage

Efficiency & Scale

CategoryDescriptionMethods
Compression / StreamingLightweight, mobile, and progressive streamingCompact-3DGS, LightGS, MobileGS, Embedded-3DGS, NanoGS, OT-UVVS, Gaussians on a Diet, HAC, MesonGS++, CAGS, PD-4DGS, MGS, DSGS, VEDAL
AccelerationTraining and inference speedupProxy-GS, Faster-GS, GEMM-GS, AV1-3DGS, BlitzGS, 3DGS³, TensorGS, HiGS, DSD-GS
Large-ScaleCity-scale and distributed scene managementBlitzGS
Feed-ForwardSingle-forward-pass generalizable reconstructionZ-Order GS, RoSplat, SplatWeaver, AdaptSplat, PointForward, ZPressor, VolSplat, PM-Loss, DéjàView, VG²GT

Understanding & Semantics

CategoryDescriptionMethods
Language / SemanticOpen-vocabulary 3D understanding and language fieldsLangSplat, Feature 3DGS, Semantic Foam, NG-GS, ReferSplat, SCOUP, Gaga
Generation / Text-to-3DText/condition-driven 3D generationDreamGaussian, SceneGen-LLMRL, DeG, PanoWorld, AnySurf, AssetGen, SEIG, MORPHOS
Autonomous DrivingDriving scene reconstruction and simulationReal2Sim, ConFixGS, P2GS, GEM, StreetNVS

Dynamic & Spatial

CategoryDescriptionMethods
Dynamic4D Gaussians, temporal deformation, and propagationParticleGS, 3DGS³, Velox, RetroNVS, WebSpline
HDR / DynamicHDR capture and time-varying illuminationHDR-NSFF, FreeTimeGS++
SLAMSimultaneous localization and mapping2DGS-SLAM, MAGS-SLAM, ULF-Loc, LiteLoc, Flow4DGS-SLAM, GGD-SLAM
Sparse-ViewFew-shot and sparse-view reconstructionFrameTwin, GeoQuery, VidSplat, PanoPlane

Applications & Cross-Domain

CategoryDescriptionMethods
Human / AvatarAnimatable human and avatar reconstructionGaussianAvatar, GAS, SplattingAvatar, Generalizable Human GS, HumanSplatHMR, D-Rex, DelightingFace, HairGPT, ArtMesh, HeadsUp, CapTalk, 3DGS Head Avatars
EditingInteractive and text-guided scene editingGaussianEditor, GeoGaussian, Frosting, SketchFaceGS, FluSplat, TransSplat, SVGS (Edit), VIRGi, GOR-IS, FaceParts, Robust Prior-Guided Seg., AlbedoEdit
RelightingRelightable and material-aware GaussiansRelightable-GS-VP, Ambient-Robust IR
CADCAD model fitting and reverse engineeringCADFit, KDH-CAD
Cross-DomainMulti-modal fusion and out-of-domain transferGS-DOT, DiffSoup, FTSplat, IRIS, SplAttN, Fake3DGS, RGS, RESPIRE, LagrangianSplats, PG-3DGS, OCH3R, TransmissiveGS, SurfFill
SimulationPhysics simulation and surrogate modelsGS-Playground, GS-Surrogate, FieryGS, SAM3D-Phys
Embodied AI / RoboticsRobotic grasping, manipulation, navigation, and sim-to-realGaussianGrasper, GraspSplats, ManiGaussian, GSMem, RoboSplat, VR-Robo, GSDrive, LEGS
Active VisionActive view selection and next-best-viewMAGICIAN
Real-Time NVSReal-time novel view synthesis systems3DTV
Data AcquisitionCapture protocols and phone-based acquisitionMobile Phone 3DGS Acquisition
RobustnessIn-the-wild and degradation-robust reconstructionNRGS, DualSplat, EnerGS, FreeFix, Luminance-GS++, HarmoGS
Degradation-AwareWeather, underwater, and low-quality input handlingMERID-GS, MarineSTD-GS, E2EGS, NIRRGB-GS
SystemFull-system design and hardware co-designYOGO, GS-SCNet
SecurityWatermarking, copyright protection, and forgery detectionRDSplat, GuardMarkGS, 3DEditSafe
World Model3DGS world models and scene predictionMRO-GWM
Event CameraEvent camera-driven 3DGSTIDES

Quick Start

Each skill is a standalone SKILL.md file — copy it to your Agent's skills directory and it works immediately. No dependencies, no build steps.

3 commands to your first AI-powered 3DGS workflow:

git clone https://github.com/jaccen/Awesome-Gaussian-Skills.git
cp -r Awesome-Gaussian-Skills/skills/* ~/.openclaw/skills/   # or .claude/ for Claude Code
openclaw restart   # then ask: "Compare 3DGS and 2DGS rendering formulations"

Option 1: OpenClaw

# Copy all skills to OpenClaw skills directory
cp -r skills/* ~/.openclaw/skills/
openclaw restart

Option 2: Claude Code / Cursor

# Copy the skills you need into your project's .claude/ directory
cp skills/3dgs-paper-reader/SKILL.md .claude/
cp skills/3dgs-code-reviewer/SKILL.md .claude/

Option 3: One-Click Install Script

curl -sSL https://raw.githubusercontent.com/jaccen/Awesome-Gaussian-Skills/main/scripts/setup.sh | bash

Skills Overview

1. 3dgs-paper-reader — Paper Reading & Summarization

Read any 3DGS paper and extract structured insights in seconds.

You: "帮我读一下这篇论文 2401.01345,总结核心方法和实验结果"
Agent: [Extracts method name, core innovation, architecture, 
        training strategy, datasets, metrics, and key findings]

Capabilities:

  • Parse arXiv papers (by ID or URL) and local PDFs
  • Extract method architecture, loss functions, training details
  • Generate structured summaries with tables
  • Identify novelty and limitations

2. 3dgs-method-compare — Method Comparison Engine

Compare any 3DGS variants across 10+ dimensions.

You: "对比 ***和 **** 的核心差异"
Agent: [Generates comparison table across: primitive representation,
        opacity handling, color mechanism, frequency modeling,
        geometric boundary, training strategy, etc.]

Capabilities:

  • Multi-method side-by-side comparison
  • 10+ comparison dimensions (rendering formula, primitive, loss, speed, ...)
  • Identify design trade-offs
  • Built-in knowledge of 660+ methods

3. 3dgs-code-reviewer — Implementation Code Review

Catch common 3DGS implementation bugs before submission.

You: "审查我的 3DGS CUDA 渲染 kernel,看看有没有性能问题"
Agent: [Checks: alpha compositing order, tile-based rendering,
        CUDA memory coalescing, gradient computation, etc.]

Capabilities:

  • Review CUDA kernels for correctness and performance
  • Check rendering pipeline (differentiable rasterization, alpha blending)
  • Validate loss function implementations
  • Common bug pattern detection (97+ known patterns)

4. 3dgs-experiment-planner — Experiment Design Assistant

Design rigorous experiments that satisfy top-venue reviewers.

You: "我要写一篇关于高频边界建模的3DGS论文,帮我设计实验"
Agent: [Recommends datasets, baselines, metrics, ablation matrix,
        rendering comparisons, and efficiency analysis plan]

Capabilities:

  • Dataset and baseline recommendation (tailored to your method)
  • Ablation study matrix design
  • Metric selection guidance (PSNR/SSIM/LPIPS/FID/CHF/etc.)
  • Efficiency analysis framework
  • Visualization plan for paper figures

5. nerf-to-3dgs-migrator — NeRF-to-3DGS Migration Guide

Migrate your NeRF method to 3DGS with step-by-step guidance.

You: "我的 NeRF 方法用了 hash encoding + deformable field,
     怎么迁移到 3DGS?"
Agent: [Provides migration plan: feature encoding → per-Gaussian 
        attributes, deformation → position/rotation offsets, 
        with code templates]

Capabilities:

  • Component-level migration analysis
  • Code templates for common migration patterns
  • Identify incompatibilities and workarounds
  • Performance comparison estimation

6. cad-mesh-3dgs — CAD, Mesh & 3DGS Bridge

Navigate the mesh↔3DGS pipeline, CAD reverse engineering, and surface extraction.

You: "我训练了一个3DGS模型,怎么提取高质量的mesh?"
Agent: [Recommends SuGaR or 2DGS pipeline, provides TSDF extraction
        steps, Marching Cubes parameters, and quality evaluation code]
You: "如何把CAD模型(STEP格式)转换为3DGS表示?"
Agent: [Provides mesh→Gaussian conversion pipeline, covariance 
        initialization from mesh normals, and curvature-aware sampling]

Capabilities:

  • Mesh→3DGS conversion (sampling, initialization, optimization)
  • 3DGS→Mesh extraction (SuGaR, 2DGS, TSDF+Marching Cubes)
  • CAD reverse engineering pipeline (mesh→B-rep via primitive fitting)
  • Hybrid representation analysis (MaGS, UniMGS, 2DGS)
  • Geometry quality evaluation (Chamfer Distance, F-Score, Normal Consistency)
  • Debugging common mesh-Gaussian conversion issues

7. cg-paper-writing — CG Paper Writing Assistant v2.0

Write publication-ready papers for CVPR/ICCV/ECCV/SIGGRAPH/TVCG, with multi-agent adversarial review and citation integrity gates.

You: "帮我写一篇关于 3DGS的论文引言,要和 ****GS 做对比"
Agent: [Generates academic introduction with proper structure,
        domain terminology, and argumentation flow]

Capabilities:

  • Venue-specific writing conventions (CVPR vs SIGGRAPH vs TVCG)
  • Domain terminology database (3DGS, NeRF, rendering, geometry)
  • De-AI-ification (remove AI writing patterns)
  • Section-by-section writing (Abstract → Introduction → Related Work → Method → Experiments → Conclusion)
  • Mathematical notation conventions
  • v2.0: Multi-agent adversarial review with Concession Threshold Protocol (anti-sycophancy)
  • v2.0: Citation three-layer verification (existence, claim-faithfulness, venue currency)
  • v2.0: Integrity Gates (Post-Draft Gate + Pre-Submission Gate, non-skippable)
  • v2.0: Style Calibration (learn from user's past writing samples)
  • v2.0: Persistent writing context across sessions (symbol table, ref registry, gate status)

8. 3dgs-visualizer — Research Visualization

Generate publication-quality charts for 3DGS research: radar charts, comparison tables, and method timelines.

You: "画一个雷达图对比 3DGS、2DGS 和 NegGS 在各维度的表现"
Agent: [Generates radar chart with 7 dimensions: Render Quality,
        Speed, Memory, Geometry, Scalability, Ease of Use, Novelty]
You: "生成3DGS领域的时间线演进图,从2023年到2026年"
Agent: [Creates chronological timeline showing 40+ methods across
        25 categories with citation-weighted node sizing]

Capabilities:

  • Radar charts for multi-dimensional method comparison (7 default dimensions, customizable)
  • Visual performance/efficiency comparison tables with color-coded highlighting
  • Method evolution timelines with category lanes and citation-weighted sizing
  • Dual output: static (PDF/PNG via matplotlib/seaborn) and interactive (HTML via plotly)
  • 3 pre-built presets: Landscape Overview, Category Deep Dive, Paper Submission Package
  • Publication-quality styling with Okabe-Ito colorblind-safe palette

9. 3dgs-engineering-guide — Engineering Deployment Guide

Bridge the gap from academic 3DGS research to production deployment.

You: "How do I deploy 3DGS for autonomous driving simulation?"
Agent: [Recommends aiSim 6 / CARLA+3DGS pipeline, maps relevant
        papers (GSDrive, GS-Playground), highlights sensor
        simulation fidelity requirements]

Capabilities:

  • 10 industry application tracks (AD, digital twin, heritage, film/game, e-commerce, inspection, AR/VR, BIM, robotics, military)
  • 5-layer engineering technology stack (acquisition, reconstruction, post-processing, deployment, integration)
  • Tool selection decision trees by use case, platform, and scene scale
  • Cross-platform deployment guide (CUDA, Vulkan, WebGPU, mobile)
  • Engineering best practices and common pitfalls

10. patent-software-ip — Patent & Software Copyright Generation

Generate patent application docs and software copyright registration materials from your research project.

You: "帮我的3DGS大场景重建项目生成专利申请文件"
Agent: [Scans project code & docs, searches prior art,
        generates claims, specification, and abstract
        with desensitization and self-check]
You: "我要给这个工具申请软件著作权,帮我准备材料"
Agent: [Generates software manual (10-15 pages with screenshots)
        and source code document (front/back 30 pages each),
        formatted for CPCC registration]

Capabilities:

  • Dual-path workflow: patent application (claims + specification + abstract) and software copyright (manual + source code doc)
  • Prior art search and differentiation analysis
  • Claims drafting with independent + dependent claims structure
  • Specification with full implementation details and desensitization
  • Software copyright manual following CPCC format (10-15 pages, 6+ screenshots)
  • Source code document formatting (50+ lines/page, sensitive data removal)
  • Internal self-check for consistency, adequacy, and compliance

11. 3dgs-spatial-agent — 3DGS/CAD/Mesh Spatial Intelligence Agent

Agent-driven 3D scene reasoning, CAD extraction, and interactive editing.

You: "帮我从3DGS模型中提取椅子的CAD参数化模型"
Agent: [Segments chair Gaussians via OP2GS, extracts mesh via SuGaR,
        fits parametric primitives via GS-CAD, exports STEP file]

Capabilities:

  • Scene-level reasoning: segment → extract geometry → infer materials → build scene graph
  • CAD-in-the-loop: build123d/Open Cascade for parametric model extraction from 3DGS
  • Multi-modal I/O: text/prompt → parameterized CAD model or 3DGS scene edits
  • Articulation discovery: identify articulated structure from Gaussian grouping
  • Material inference: PBR properties from SH coefficients and Gaussian density

12. 3dgs-mcp-renderer — MCP Protocol + 3DGS Rendering Bridge

Agent-controlled Three.js/WebGPU rendering via MCP protocol.

You: "从上方看这个场景"
Agent: [Calls MCP set_camera(0,10,0), render_frame(), returns image]

Capabilities:

  • MCP tools: import_scene, set_camera, modify_gaussians, render_frame, query_scene
  • Voice-driven reconstruction: Whisper → Agent → MCP → rendering
  • Real-time parameter manipulation: opacity, color, position, scale
  • Semantic querying: label-based Gaussian selection (OP2GS, Gaga integration)

Architecture

Awesome-Gaussian-Skills/
├── skills/
│   ├── 3dgs-paper-reader/       # Paper reading & summarization
│   │   └── SKILL.md
│   ├── 3dgs-method-compare/     # Method comparison engine
│   │   └── SKILL.md
│   ├── 3dgs-code-reviewer/      # Code review for 3DGS implementations
│   │   └── SKILL.md
│   ├── 3dgs-experiment-planner/ # Experiment design assistant
│   │   └── SKILL.md
│   ├── nerf-to-3dgs-migrator/  # NeRF-to-3DGS migration guide
│   │   └── SKILL.md
│   ├── cad-mesh-3dgs/          # CAD, Mesh & 3DGS bridge (+ build123d pipeline)
│   │   └── SKILL.md
│   ├── 3dgs-visualizer/        # Research visualization (radar, table, timeline)
│   │   └── SKILL.md
│   ├── cg-paper-writing/        # CG paper writing assistant
│   │   └── SKILL.md
│   ├── 3dgs-engineering-guide/   # Engineering deployment guide
│   │   └── SKILL.md
│   └── patent-software-ip/       # Patent & software copyright generation
│       └── SKILL.md
│   ├── 3dgs-spatial-agent/       # 3DGS/CAD/Mesh spatial intelligence agent
│   │   └── SKILL.md
│   └── 3dgs-mcp-renderer/        # MCP protocol + Three.js/3DGS rendering bridge
│       └── SKILL.md
│   │   └── SKILL.md
├── scripts/
│   ├── setup.sh                 # Quick install script
│   ├── cad2gs_pipeline.py       # CAD → 3DGS conversion (STEP/GLB → Gaussian init)
│   └── build_part_aware_scenes.py # Part-Aware experiment scene builder
├── docs/
│   ├── index.html               # Interactive method explorer
│   ├── 3dgs-viewer.js           # Sidecar-powered 3DGS visualization viewer
│   ├── sidecars/                # Animation sidecar modules
│   │   ├── 3dgs-render-process.step.js  # Rendering process animation
│   │   └── 3dgs-method-compare.step.js  # Method comparison animation
│   ├── fusion-framework.md      # Fusion design principles & roadmap
│   └── fusion-demo.html         # Fusion roadmap visualization
├── Text2Word/                # Interactive text-to-3DGS web demo
│   └── index.html
├── references/
|   ├── 3dgs-methods-overview.md # Index (660+ methods across 25 categories)
│   ├── methods-core.md         # Core methods (Foundation→Dynamic)
│   ├── methods-semantic-editing.md # Semantic, Editing, Material, Avatar
│   ├── methods-systems-apps.md # Systems, Applications, Cross-Domain
│   ├── cad-3d.md               # CAD/3D terminology, baselines, build123d pipeline
│   ├── experiments.md          # Experiment design, CAD benchmark scenes
│   └── baselines.md            # Baseline methods & datasets
├── Test/
│   ├── radar_comparison.pdf/png/html       # Radar chart: 3DGS vs 2DGS vs Mip-Splatting vs Scaffold-GS
│   ├── metrics_bar_comparison.pdf/png      # PSNR / SSIM / LPIPS grouped bar chart
│   ├── quality_vs_speed_scatter.pdf/png    # Quality vs Speed bubble chart
│   ├── metrics_heatmap.pdf/png             # Normalized metrics heatmap
│   └── metrics_dashboard.html              # Interactive 4-in-1 dashboard (plotly)
├── README.md
├── README_CN.md
├── CONTRIBUTING.md
└── LICENSE

Each skill follows the SKILL.md standard (YAML frontmatter + Markdown instructions), compatible with:

  • OpenClaw (ClawHub ecosystem)
  • Claude Code (.claude/ directory)
  • Cursor (.cursor/rules/)
  • Windsurf and other AI Agent frameworks

Visualization Samples

Samples generated by 3dgs-visualizer — see Test/ for full-resolution files.

Radar Chart (Method Comparison)Metrics Bar Chart (PSNR/SSIM/LPIPS)
Quality vs Speed ScatterNormalized Metrics Heatmap

Interactive versions (hover for details): radar_comparison.html | metrics_dashboard.html

Research Innovation Highlights

Derived from systematic gap analysis across 660+ methods. Each idea identifies a concrete problem, a methodological approach, and an implementation path. Target venues: TVCG / CGF / CAD / T-RO / IJCV / ACM TOG / Pattern Recognition.

Problem: Standard alpha-compositing causes color bleeding at part boundaries of articulated objects. ULF-Loc (CVPR 2026) exposed this feature bias, but no rendering-formulation-level fix exists for the articulated case.

Approach: Extend alpha-compositing with part-aware opacity modulation: C(theta) = sum_i T_i * alpha_i * omega_{p(i)}(theta) * c_i(theta), where omega penalizes penetration and joint violations, making inter-part penetration regions automatically transparent.

Path: 1) Build on gsplat rasterizer. 2) Add FK layer for articulated objects (URDF). 3) Compute penetration/joint violation via SDF. 4) Train on Articulate-100 (ArtMesh, CVPR 2026). 5) Evaluate: NVPSNR + physical consistency metrics.

Target: SIGGRAPH / ACM TOG / TVCG

Problem: Flow-based world models (RoboFlow4D, 2026) predict dense 3D flows but lack geometric consistency -- predicted flows can violate object rigidity and physical constraints.

Approach: Couple 3D flow prediction with scene graph constraints: rigidity loss for static objects, articulation loss for joints, support-relation constraints for stacking. Slow-fast collaboration between flow prediction and scene-graph updates.

Path: 1) Extend RoboFlow4D architecture. 2) Scene graph parser via OpenMask3D. 3) Geometric regularizer (rigid/articulated losses). 4) Train on LIBERO + RoboCasa. 5) Baselines: RoboFlow4D, RISE, DreamerV3.

Target: IJCV / T-RO / RSS

Problem: Cross-modal distillation for 3D detection suffers from representation collapse when transferring image features to point clouds. Euclidean space is inherently limited for hierarchical object relationships.

Approach: Conduct distillation in hyperbolic space (Poincare ball). Leverage exponentially increasing volume to better preserve semantic hierarchy during image-to-point-cloud fusion. Three modules: SGVO (semantic-guided voxel optimization), HFT (hyperbolic feature transfer), FAGO (feature aggregation optimization).

Path: 1) Hyperbolic embedding layers (PoincareBall ops in PyTorch). 2) Dual-branch: Swin-B + VoxelNet. 3) Hyperbolic contrastive loss. 4) Datasets: SUN RGB-D, nuScenes. 5) Metrics: mAP, NDS.

Target: IEEE T-MM / T-IP / Pattern Recognition

Problem: VLMs fail on solid geometry (3D volume, cross-sections). Hilbert-Geo (CVPR 2026) introduces formal language but lacks verification guarantees.

Approach: Neural parser + Z3/SMT formal verifier in iterative refinement loop. VLM proposes reasoning steps, formal system verifies each step, incorrect steps trigger re-generation. Extends Hilbert-Geo's CDL with full 3D geometry predicates (volume, unfolding, etc.).

Path: 1) Extend CDL predicate library. 2) Train parser on SolidFGeo2k + expanded synthetic data. 3) Integrate Z3 Python API. 4) Evaluate: accuracy + formal correctness rate. 5) Baselines: Hilbert-Geo, GPT-4o, Gemini.

Target: Pattern Recognition / Neural Networks / AAAI / ICLR

Problem: GSMem uses 3DGS as spatial memory but lacks semantic persistence -- object semantics degrade as the agent explores new areas.

Approach: Hippocampus-inspired memory: short-term 3DGS scene graph + periodic semantic consolidation into compact embeddings (Perceiver compressor) + importance-weighted forgetting + retrieval-augmented long-horizon planning.

Path: 1) Base: GSMem + SceneGPT. 2) Memory consolidation: Perceiver-based compressor. 3) Episodic replay for forgotten regions. 4) Tasks: HM3D navigation, object search. 5) Metrics: SPL, semantic recall over time.

Target: T-RO / IJCV / ECCV

Problem: OrbiSim (2026) demonstrates differentiable physics potential but lacks realistic contact modeling for manipulation (grasping, pushing, stacking).

Approach: Extend OrbiSim with SDF-based contact detection for arbitrary meshes + differentiable KKT contact solver + Coulomb friction projection. Enables gradient-based policy optimization under sparse-reward contact-rich tasks.

Path: 1) Extend OrbiSim codebase. 2) Neural SDF for contact geometry (DMTet). 3) Differentiable KKT contact layer (DiffCo). 4) Tasks: MimicGen, LIBERO. 5) Baselines: OrbiSim, DreamerV3, PPO.

Target: ACM TOG / SIGGRAPH / TVCG / RSS

Problem: VLA models (RT-2, GR00T, Pi-0) rely almost exclusively on vision. The tactile survey (2026) identifies contact geometry as essential but poorly integrated.

Approach: Unified tactile-visual spatial representation: tactile images -> contact geometry maps (normal, depth, shear) -> project onto 3D scene -> cross-attention fusion transformer -> shared action decoder.

Path: 1) Tactile simulation: GelSight in IsaacSim. 2) Visual backbone: DINOv2. 3) Differentiable tactile-to-3D projection. 4) Training: insertion, assembly tasks. 5) Datasets: MIT Touch, Slippery manipulation.

Target: T-RO / IJCV / ICRA

Problem: PanoWorld generates panoramic video as visual synthesis. NavSpace tests spatial instruction following but lacks generative prediction. Neither couples visual generation with spatial affordances for navigation.

Approach: Joint panoramic world model: spherical visual panorama + top-down semantic BEV with affordances + traversability segmentation + spatial relation graph. Enables "imagine then navigate" planning loop.

Path: 1) Adapt PanoWorld's spherical diffusion model. 2) Add parallel BEV semantic decoder. 3) Affordance head (traversability + interaction). 4) Training: Matterport3D, HM3D. 5) Integration: plug into NavSpace for closed-loop nav.

Target: ECCV / IJCV / CVPR

Problem: SpatialBabel (2026) reveals VLMs can generate correct 3D code but fail at simpler spatial questions -- code serves as external memory the model cannot access internally.

Approach: Self-supervised pipeline: VLM generates Three.js code -> execute in headless renderer -> extract spatial annotations (depth, normals, relations) -> fine-tune VLM on extracted QA pairs -> at test time, answer without code generation.

Path: 1) Extend S3-FT from SpatialBabel. 2) Automated: prompt -> code -> render -> annotation. 3) 10K+ procedural 3D scenes. 4) Fine-tune Qwen3-VL. 5) Eval: SpatialBabel, CV-Bench-2D/3D/Spatial.

Target: CVPR / NeurIPS / AAAI

Problem: Occupancy (SparseWorld, DOV) is the driving world model standard; 3DGS provides superior rendering. No differentiable bridge exists between them.

Approach: Bidirectional conversion: Occ->3DGS (differentiable sampler from occupancy+semantic features to Gaussian parameters) and 3DGS->Occ (learned voxel pooling from Gaussian parameters). Joint backbone for unified prediction+rendering.

Path: 1) Backbone: SparseWorld-TC or DOV. 2) Occ->3DGS: learned position+scale predictor. 3) 3DGS->Occ: differentiable sparse convolution pooling. 4) Training: nuScenes, Waymo. 5) Metrics: mIoU, PSNR, latency.

Target: TVCG / T-ITS / CVPR / ICCV

Roadmap

  • v0.1 — Initial release with 6 core skills (Apr 2026)
  • v0.1.1 — Add cad-mesh-3dgs skill for CAD/Mesh↔3DGS bridge (Apr 2026)
  • v0.1.2 — Knowledge base expansion: 50→120+ methods, 23 categories, daily auto-update workflow (Apr-May 2026)
  • v0.1.3 — Knowledge base v2: 130→150+ methods, 52+ bug patterns, 23 categories, cross-domain expansion (May 2026)
  • v0.1.4 — Knowledge base v3: 150→152+ methods, 53+ bug patterns, 23 categories (optimized), FreeTimeGS++, D-Rex (May 2026)
  • v0.1.5 — Knowledge base restructured: split overview into 3 sub-files for efficient retrieval (May 2026)
  • v0.2 — Add 3dgs-visualizer skill (radar charts, comparison tables, method timelines; static + interactive output) (May 2026)
  • v0.2.1 — Add Text2Word interactive demo — text-to-3DGS web creation platform (May 2026)
  • v0.2.2 — Add patent-software-ip skill — patent application docs & software copyright registration materials (May 2026)
  • v0.2.3 — Knowledge base expansion: 222→240+ methods, 57+ bug patterns, daily update + new physics-GS/streaming methods (May 2026)
  • v0.2.4 — Knowledge base expansion: 240→254+ methods, 60+ bug patterns, OIT transparency, distributed city-scale GS, 4DGS physics sim for AD (May 2026)
  • v0.2.5 — Knowledge base expansion: 254→512+ methods, 62+ bug patterns, Skew-Normal primitive, expert-routing feedforward GS, continuous LoD, adaptive split operator (May 2026)
  • v0.2.6 — Daily update: 512→516+ methods, +4 new (Learn2Splat meta-optimizer, EndoGSim medical 4DGS+MPM, 3DEditSafe editing safety, Robust Prior-Guided Segmentation); Medical imaging expanded (May 2026)
  • v0.2.7 — Daily update: 516→531+ methods, +15 new (HarmoGS, ULF-Loc, AdaptSplat, PointForward, 3DGS³, TransmissiveGS, NIRRGB-GS, TensorGS Tensor Core, DeG SIGGRAPH'26, P2GS CVPR'26, ArtMesh, DSGS, GEM, PanoWorld, LiteLoc); 63+ bug patterns; 25 categories (May 2026)
  • v0.2.8 — Daily update: 531→548+ methods, +17 new (ZPressor NeurIPS'25, VolSplat, PM-Loss 3DV'26, AmbiSuR ICML'26, RT-Splatting CVPR'26 Highlight, TideGS 1B+ Gaussians, OP2GS dual-opacity, MMGS OT compression, 3DSGS skew Gaussian, GaussianZoom zoom-in gen, AnyCity aerial, Cross-View Splatter, FLUIDSPLAT, GS-DIFF, ReorgGS, AsyncEvGS, SplitGS-Loc); 67+ bug patterns; 25 categories (May 2026)
  • v0.2.9 — CAD fusion: cad-mesh-3dgs skill + build123d pipeline (STEP→GLB→3DGS), cad2gs_pipeline.py, Part-Aware experiment scene builder, Sidecar animation system for 3DGS visualization (May 2026)
  • v0.2.10 — Daily update: 548→559+ methods, +5 new (GaussianPile CVPR'26 volumetric medical 3DGS, Flow4DGS-SLAM CVPR'26 flow-guided dynamic SLAM, Ilov3Splat ICPR'26 instance-level open-vocab, LeGS learnable density control, PhysX-Omni unified physical 3D gen); 69+ bug patterns; 25 categories (May 2026)
  • v0.2.11 — Daily update: 553→559+ methods, +6 new (CAdam SIGGRAPH'26 generative densification, GGD-SLAM ICRA'26 dynamic SLAM, PiG-Avatar volumetric canonical avatars, Latent Dynamics clothing animation, ROAR-3D multi-view generation, GLUT color transformation) (May 2026)
  • v0.2.12 — Daily update: 559→574+ methods, +15 new (ArtSplat first feed-forward articulated 3DGS 400x faster, BEA-GS CVPR'26 Highlight object extraction, F-RNG feed-forward relightable 25x faster, CodecSplat 20-108 KiB/scene, TriSplat triangle primitives mesh export, VoxelGS scaffold-anchored Gaussian+voxel SDF, TokenGS learnable tokens, NoPo4D pose-free feed-forward 4DGS, COSY compositional head editing, R5DGS semantic 4DGS rigid constraints, RiGS rigid-aware monocular 4DGS, Underwater360/GlowGS/DelowlightSplat degradation-specific 3DGS, RoVES physics-aware driving editing, 4D-GSW 4DGS watermarking, DP-GES depth peeling surfel rendering); 71+ bug patterns; 25 categories (May 2026)
  • v0.2.13 — Daily update: 574→591+ methods, +17 new (EulerianGS CVPR'26, GAVIS CVPR'26, DGSG-Mind, DeGO CVPR'26, TWINGS CVPR'26, BitC-3DGS, FRUC, IDESplat CVPR'26, PhyGenHOI, MonoPhysics, FreeForm CVPR'26, TDg, SRUG, NeuROK CVPR'26, PocketGS, X-GS, WeatherCity); 79+ bug patterns; 25 categories (May 2026)
  • v0.2.14 — Daily update: 591→607+ methods, +16 new (HeadsUp Apple UV-parameterized head, DéjàView looped transformer, View-Dependent Splatting Kernels SIGGRAPH'26, SAM3D-Phys generative priors+physics, Gaga 3D-aware memory bank, CapTalk text+speech head, SurfFill LiDAR+surfel, CLEAR-NeRF unbounded, AnySurf/AssetGen/DinoComplete generation+completion, Ambient-Robust IR RGB-NIR, Multi-view Consistent 3DGS Head CVPR'26, Gaussian-Enhanced Surfel depth-peeling); 82+ bug patterns; 25 categories (Jun 2026)
  • v0.3.0 — Skills standard alignment (Anthropic Claude Code / OpenClaw) + Daily update: 607→630+ methods, +23 new (HiGS NVIDIA 15.8x rendering, DDF-GS ray queries, VEDAL variational pruning, StreetNVS multi-sensor NVS, LEGS embodied GS sim, KDH-CAD knowledge-hybrid, SEIG VLM→Blender, MORPHOS 4D gen, WebSpline spline trajectories, AlbedoEdit video editing, MRO-GWM world model, DSD-GS 700FPS, VG²GT voxel-gaussian, TIDES event sim, Triangle Splatting SLAM, DeblurNVS, GeoSAM-3D, SplatShot avatar, GSDeformer deformation, MidSurfNet mid-surface, 3DCodeBench benchmark, Dynamic Mesh-Gaussian physics); 88+ bug patterns; 25 categories (Jun 2026)
  • v0.3.1 — Daily update: 630→660+ methods, +18 new (ZipSplat token-based feed-forward, Geometry Gaussians per-splat geometry opacity, 3DReflecNet CVPR'26 reflective benchmark, MeshFlow CVPR'26 Highlight MeshVAE+Diffusion, MeshWeaver CVPR'26 autoregressive mesh, UniCAD multi-task CAD benchmark, Anchor3R streaming reconstruction, SimuScene physics-in-the-loop, 3DThinkVLA latent 3D priors, SymTRELLIS symmetry-aware generation, HSP/GemNR/T2Mo/LetCamsGo/MetaPoint/PureLight + 2 surveys); 97+ bug patterns (#89-#93); Skills v1.6-1.9 across 6 updated skills; Token-based cluster in method comparison (Jun 2026)
  • v0.3.2 — Daily update: 648→660+ methods, +12 new (RAF, FreeForm, D4RT CVPR Best Paper, TRELLIS.2 Best Student Paper, SAM 3D Honorable Mention, ReLaGS, FreeArtGS, ArtGS, PARTICULATE, DropAnSH-GS, BA-GS, SR3R); 97+ bug patterns (#94-#97); 8 Skills updated; MCP tools +3 (simulate_physics, query_4d_scene, deform_elastic) (Jun 2026)
  • v0.4 — Advanced 3dgs-spatial-agent skill enhancements (knowledge-constrained CAD via KDH-CAD, DDF-GS ray query, SEIG procedural generation) + SignedGS/TPSC domain skill
  • v0.5 — MCP protocol integration: Agent-controlled Three.js/3DGS rendering pipeline + DDF-GS ray query MCP tool
  • v1.0 — ClawHub official listing + CI/CD integration + anthropics/skills PR
  • v1.1 — Multi-language support (Chinese, Japanese, Korean)
  • v2.0 — Agent-to-Agent collaboration (multi-agent paper discussion)

Contributing

Contributions are welcome! Please read our Contributing Guide for details.

Ways to contribute:

  • Add new skills for uncovered scenarios
  • Expand the methods knowledge base
  • Report bugs or suggest improvements
  • Share your use cases and success stories

Contributors

Citation

If you find this project helpful in your research, please consider citing:

@misc{awesome-gaussian-skills,
  author = {jaccen},
  title = {Awesome Gaussian Skills: 3D Spatial Intelligence Open-Source Toolbox for 3D Gaussian Splatting Research},
  year = {2026},
  url = {https://github.com/jaccen/Awesome-Gaussian-Skills}
}

Acknowledgments

License

Apache-2.0. See LICENSE for details. Fork, publish, sell — with explicit patent grant for enterprise users.

Star History

Star History Chart

If this project saves you time, please give it a star!

Ver o original no GitHub

Bom para

Relacionados